Skip to content

config.md

Latest commit

 

History

History
196 lines (142 loc) · 5.02 KB

config.md

File metadata and controls

196 lines (142 loc) · 5.02 KB

Config Reference (~/.loopforge/config.toml)

LoopForge stores configuration in ~/.loopforge/config.toml (path kept for compatibility).

Mental model

Think of the file in four layers:

  1. providers.* — how to talk to each model provider
  2. router.* — which provider/model to use for planning, coding, and summary work
  3. security.* — what network, secret, and leak-guard rules apply around tool execution
  4. skills.* — how local skills are allowlisted and approved

In practice:

  • loopforge init creates the baseline config
  • loopforge config validate checks whether the file parses and the required structure is present
  • loopforge doctor helps explain why a config is valid-but-still-not-usable in your environment

MCP servers (per session)

MCP servers are not persisted in config.toml yet. Enable them per run:

loopforge agent run \
  --workspace my-ws \
  --mcp-config mcp-servers.json \
  --prompt "…"

Validate and inspect

loopforge config validate
loopforge config validate --json
loopforge doctor

Use config validate for syntax and schema issues. Use doctor for runtime readiness issues such as missing env vars, browser prerequisites, or security posture warnings.

Minimal starter example

[providers.ollama]
kind = "openai_compatible"
base_url = "http://127.0.0.1:11434/v1"
api_key_env = ""
default_model = "qwen3:4b"

[router.planning]
provider = "ollama"
model = "default"

[router.coding]
provider = "ollama"
model = "default"

[router.summary]
provider = "ollama"
model = "default"

[security.secrets]
mode = "env_first"

[security.leaks]
mode = "warn"

[skills]
auto_approve_readonly = true

This is enough for a local-first Ollama setup. You can harden it later by adding security.egress.rules and tightening skills policy.

Providers

Each provider entry defines:

  • kind: driver kind (openai_compatible, zhipu_native, minimax_native, bedrock, etc.)
  • base_url: API base URL
  • api_key_env: environment variable name that contains the API key (empty for local providers)
  • default_model: default model for model = "default"

Example:

[providers.ollama]
kind = "openai_compatible"
base_url = "http://127.0.0.1:11434/v1"
api_key_env = ""
default_model = "qwen3:4b"

AWS Bedrock (Converse API)

For AWS Bedrock, use kind = "bedrock" plus an aws_bedrock table:

[providers.bedrock]
kind = "bedrock"
base_url = ""     # unused for Bedrock
api_key_env = ""  # unused for Bedrock
default_model = "anthropic.claude-3-5-sonnet-20241022-v2:0"

[providers.bedrock.aws_bedrock]
region = "us-east-1"
cross_region = "" # optional
profile = ""      # optional

Notes:

  • Bedrock uses the AWS SDK credential chain (env vars, shared config, profiles, instance role, etc.).
  • cross_region (optional) prefixes model ids with <cross_region>. when they are not already prefixed.

Router

Each task kind selects a (provider, model) pair. This is how the runtime decides whether planning, coding, and summary turns should use the same model or different ones:

[router.planning]
provider = "ollama"
model = "default"

[router.coding]
provider = "ollama"
model = "default"

[router.summary]
provider = "ollama"
model = "default"

Security

[security.secrets]
mode = "env_first"

[security.leaks]
mode = "redact"

[[security.egress.rules]]
tool = "web_fetch"
host = "docs.rs"
path_prefix = "/"
methods = ["GET"]

Fields:

  • security.secrets.mode
    • env_first: resolve provider credentials from host environment variables
  • security.leaks.mode
    • off: do nothing extra
    • warn: annotate likely secret leaks but keep raw output
    • redact: mask detected ranges before persistence and follow-up model calls
    • enforce: block the tool result when likely secrets are detected
  • security.egress.rules
    • when empty, LoopForge keeps baseline SSRF/private-network guards only
    • when non-empty, outbound requests must match an allow rule in addition to baseline guards

Each egress rule contains:

  • tool: tool name, for example web_fetch
  • host: exact destination host
  • path_prefix: required URL path prefix
  • methods: allowed HTTP methods

Current outbound allowlist enforcement applies to web_fetch, A2A requests, and browser navigation entrypoints.

Built-in presets

LoopForge includes common provider presets (names may evolve):

  • OpenAI-compatible: deepseek, kimi, qwen, glm, minimax
  • Provider-native: glm_native, minimax_native, qwen_native

Skills

This section controls local skill policy. It is separate from the normal tool sandbox: a skill can still be blocked even when the underlying workspace is otherwise valid.

[skills]
allowlist = ["hello-skill", "qa-helper"]
require_approval = false
auto_approve_readonly = true
experimental = false

Fields:

  • allowlist: optional global skill allowlist
  • require_approval: force approval for non-readonly skills
  • auto_approve_readonly: when true, readonly skills skip manual approval
  • experimental: optional flag for rollout messaging